I've spent a non trivial amount of time trying to eliminate the test errors, but
it's looking like this is pretty dependent on Unix support.  For example, there
are attempts to delete open files, and uses of threads that report I/O attempts
on closed files.  (Maybe this is a race condition?  Don't we usually use
processes as workers on Windows?)

In any event, I don't want real new errors elsewhere to be masked by these known
problems.

For some reason $CACHEDIR is reported as missing in test-remotefilelog-repack.t,
but it actually exists in the hgcloneshallow call inside
shallowutil.mkstickygroupdir().  By the time the process exits, it's gone.  I
don't see it being removed by code that calls 'rmdir' or 'remove' in the
extension itself.